Is Tensor Memory Accelerator (TMA) Exposed on the DGX Spark GB10 GPU?

[n.b.agostini]

Hello,

I am looking for ways to test TMA (Tensor Memory Accelerator) performance. I am looking into the DGX Spark platform and understand that the GB10 Blackwell GPU has compute capability 12.1.

Can anyone confirm whether TMA (Tensor Memory Accelerator - introduced with hopper) is supported on SM 12.1 devices? If so, is there any public documentation describing how TMA is exposed on DGX Spark?

I am interested in using TMA-based kernels through CUTLASS/CuTe, Triton, or custom CUDA code.

I believe checking the generated SASS to be way to verify TMA support? For example:

cuobjdump --dump-sass <kernel_binary>

and looking for TMA-related instructions such as:

cp.async.bulk.tensor

Can anyone with a DGX Spark verify this is available?

I understand that I could rent H100/B100, this is my fallback plan if support for DGX Spark is not there.

Thank you in advance

[s0ne]

Consumer grade GPUs like GB10, RTX5090 and RTX6000PRO don’t have TMEM since TMEM is only existed datacenter grade GPU like B100, B200, GB200 and etc.

Inside NVIDIA DGX Spark: Is DGX Spark Actually Blackwell?